Study on deep reinforcement learning-based multi-objective path planning algorithm for inter-well connected-channels

Abstract

Defining inter-well connectivity is very important for the water injection development of carbonate fractured-vuggy reservoirs. However, most conventional methods based on logging or seismic data are subjective. In this research, multi-attribute seismic data were used to characterize the 3D search space, and an improved reward function was designed based on the law of fluid flow and appropriate heuristics. A deep reinforcement learning model based on an improved reward function was designed to search for inter-well connected channels. Selecting the shortest channel length and the largest flow volume were the optimization objectives. We propose the multi-objective evolutionary optimization algorithm with decomposition and differential evolution (MOEA/D-DE) based on an improved mutation operator to automatically obtain the optimal inter-well connected channels. The searched channels can readily show the spatial distribution of multi-scale fractures and caves. The experimental results of the Tahe oilfield show that the improved algorithm effectively enhanced the local convergence performance of the multi-objective algorithm and that the automatic search paths were consistent with the characteristics of seismic static data and tracer test results. Our model can better reflect the spatial distribution of a fracture cavity at different scales and offers important guidance for on-site development adjustment work in the water injection development stage.